figure b

Introduction

Closed-loop insulin delivery (or artificial pancreas) is, to date, the most advanced and promising technology to improve glucose control and reduce the risk of hypoglycaemia in individuals living with type 1 diabetes [1]. Closed-loop systems combine continuous subcutaneous insulin infusion (CSII), continuous glucose monitoring (CGM) and a dosing algorithm that dynamically controls the insulin infusion rate [2]. Closed-loop systems have demonstrated superior clinical efficacy (reduction in hyper- and hypoglycaemia) over conventional CSII therapy in most studies [3, 4]. Nevertheless, optimisation of closed-loop systems is challenging during exercise and meal consumption: both situations are associated with rapid changes in glucose levels and complex physiological effects [5,6,7].

During and after exercise, people with type 1 diabetes are unable to decrease their circulating plasma insulin levels in the setting of hypoglycaemia. This relative hyperinsulinaemia is exacerbated by increased insulin absorption from subcutaneous deposits that occurs in response to exercise [8]. Inappropriately elevated plasma insulin levels restrict hepatic glucose production, thus limiting counterregulatory responses to hypoglycaemia. At the same time, exercise enhances the rate of glucose disposal into skeletal muscle [9]. Therefore, physical activity is associated with a significant increase in hypoglycaemic risk in type 1 diabetes [9, 10]. Exercise is typically undertaken in a postprandial state (within 2 h of eating a meal) or later (in a post-absorptive state). Glucose management is particularly challenging during the postprandial state given the combined challenges of minimising the prandial glucose rise and mitigating the risk of exercise-induced hypoglycaemia from insulin given as a mealtime bolus [9, 11, 12].

When diabetic individuals who use CSII or multiple daily injections choose to exercise in the postprandial period, guidelines suggest a 25–75% reduction in the insulin bolus for the preceding meal [10, 11]. The degree of bolus reduction should be proportional to exercise length and intensity [11]. Those who use CSII may also choose to temporarily reduce their basal insulin infusion rate. Most available studies investigating this option have been conducted in the post-absorptive state rather than the postprandial state. Data suggest that basal insulin reduction needs to be implemented at least 40 min before exercise onset to significantly reduce circulating insulin levels given the pharmacokinetics of available insulin formulations [13]. Zaharieva et al [14] even suggest that basal rate reduction needs to be applied 90 min before exercise, to improve glucose control and decrease hypoglycaemic risk during exercise.

In the context of closed-loop systems, most exercise trials have taken place in the post-absorptive state, and postprandial exercise has not been well studied. Most published closed-loop reports to date had hybrid closed-loop systems that required some intervention by the user. Research teams have attempted to alert a closed-loop system to physical activity by adding wearable sensors to detect exercise (e.g. heart rate, energy expenditure) [15,16,17,18] or by directly announcing the start of an exercise session [6] and/or adding glucagon, a counterregulatory hyperglycaemic hormone, to the closed-loop system [4]. By announcing or detecting exercise, the closed-loop system sets a higher glucose target for its algorithm, which reduces the basal insulin infusion rate [6, 19]. Nevertheless, even with closed-loop systems, detection or announcement at the beginning or during exercise may not be enough to prevent exercise-induced hypoglycaemic risk given current insulin pharmacokinetics. For example, using closed-loop insulin delivery for a post-absorptive exercise session, announcement 20 min before the start of exercise did not fully prevent hypoglycaemia, which still occurred in 23.5% of study participants [6]. It is expected that the hypoglycaemic risk would be exacerbated in the postprandial state.

To the best of our knowledge, clinical trials specifically designed to address and compare effective strategies for postprandial exercise using closed-loop systems have not been conducted. We could identify only two in silico studies that included postprandial exercise sessions taken 2 h after a meal. The first study found that reducing basal insulin delivery by 50% 90 min before exercise and by 30% during exercise was safe and effective (meal bolus was not modified) [20]. A more recent study suggested that the combination of carbohydrate consumption with closed-loop systems is an effective strategy to mitigate the risk of exercise-induced hypoglycaemia [21].

We therefore aimed to assess the efficacy of three strategies applied in the context of a single-hormone (insulin) closed-loop system, to prevent hypoglycaemia during postprandial moderate-intensity exercise. These strategies included exercise announcement to the algorithm to increase its glucose target 90 min prior to exercise, with or without meal bolus reduction, in comparison with unannounced exercise.

Methods

Participants and study design

We conducted a single-blind, randomised, three-way crossover study with a single-hormone (insulin) closed-loop system, to compare three strategies applied during postprandial exercise in adults with type 1 diabetes: (1) announced exercise to the closed-loop system and a 33% reduction in meal bolus (A-RB); (2) announced exercise to the closed-loop system and a full meal bolus (A-FB); (3) unannounced exercise and a full meal bolus (U-FB). Exercise announcement and meal bolus reduction were applied at breakfast time, 90 min before the start of the exercise session. The study was registered at ClinicalTrials.gov (registration no. NCT02855307) and was approved by the ethics committee of the Montreal Clinical Research Institute.

Inclusion criteria were a diagnosis of type 1 diabetes for at least 1 year, age ≥18 years, use of CSII for at least 3 months and HbA1c ≤107.7 mmol/mol (≤12%). Exclusion criteria included advanced microvascular complications, a recent (<3 months) acute macrovascular event, use of medication with effects on heart rate (e.g. β-blockers), abnormal blood panel and/or anaemia, ongoing or planned pregnancy, and a severe hypoglycaemic episode within 2 weeks of screening.

Randomisation and blinding

Balanced randomisation was used to determine the order of the interventions (A-RB, A-FB or U-FB). A study coordinator conducted the randomisation and placed the results in sealed envelopes that were opened at the end of each admission visit. Participants were blinded to the sensor glucose readings and to the insulin infusion rates throughout the intervention visits.

Procedures

During the admission visit, medical data, HbA1c level, anthropometric variables and records of insulin therapy (basal and bolus doses) for the previous 3 days were collected. Physical fitness was assessed using a graded exercise test, adapted from Storer et al [22], on an ergocycle (ER 900; Ergoline, Germany) until voluntary exhaustion with power output increased by 10, 15 or 20 W/min. Expired gas samples were analysed via a mixing chamber using a Moxus cardiorespiratory test station (AEI Technologies, USA). \( \dot{V}{\mathrm{O}}_{2\mathrm{peak}} \) corresponded to the highest 30 s mean value reached.

Participants were told to avoid moderate- to high-intensity exercise the day before the intervention. A glucose sensor (Dexcom G4 Platinum; Dexcom, USA) was inserted 24 h before each intervention and calibrated by the study participant according to the manufacturer’s recommendation.

For the intervention visits, participants arrived at the testing centre at 06:30 h and stayed for 5 h and 30 min. Closed-loop control was started at 07:00 h and a standardised breakfast (65 g carbohydrates) was given at 08:00 h, with or without exercise announcement and with or without a reduction in meal bolus, according to the randomisation. The individual’s usual insulin/carbohydrate ratio was used to calculate the breakfast insulin bolus. At 09:30 h, participants performed exercise for 60 min on an ergocycle at 60% of \( \dot{V}{\mathrm{O}}_{2\mathrm{peak}} \). Exercise intensity (60% \( \dot{V}{\mathrm{O}}_{2\mathrm{peak}} \)) was monitored by indirect calorimetry (O2 and CO2) using a face mask.

Intravenous blood samples were collected every 15 min from 07:00 h to 09:30 h, every 10 min from 09:30 h to 10:30 h (exercise period) and every 15 min from 10:30 h to 12:00 h (recovery period). Blood samples were immediately processed to measure plasma glucose levels, using a YSI 2300 STAT Plus analyser (Yellow Springs, OH, USA), and were stored for subsequent measurement of insulin levels in duplicates using an immunoassay (Millipore, USA). At 11:30 h, participants were switched back to their usual CSII settings, were served a standardised meal (65 g carbohydrate) and discharged at 12:00 h. The three intervention visits were separated by a median of 7 days (IQR 7–14 days).

The closed-loop system was based on a model predictive control algorithm, as previously described [4, 6, 23]. Details may be found in the appendix of our previous study [4]. The artificial pancreas system was initialised using records of participants’ previous 3 days of insulin therapy (total daily dose, carbohydrate/insulin ratios and basal rates) and body weight obtained at the screening visit. Participants’ usual fast-acting insulin analogue was used. Exercise announcement to the algorithm increased the target glucose level from 6.0 mmol/l to 9.0 mmol/l, which was directly set back to 6.0 mmol/l at the end of exercise, as previously described [6]. Real-time sensor readings were manually entered every 10 min into the dosing algorithm running on a laptop computer. Recommendations of insulin delivery were then generated by the algorithm and applied manually through the infusion pump.

Hypoglycaemia events, necessitating glucose consumption for correction, were defined as plasma glucose <3.3 mmol/l with symptoms or <3.0 mmol/l irrespective of symptoms, as in our previous closed-loop studies [4, 6, 24]. For correction, 50 ml 20% dextrose was infused intravenously (instead of oral glucose) because of the use of facial masks for gas exchange sampling to quantify exercise intensity. Hypoglycaemia correction was repeated every 15 min until glucose levels reached 4.0 mmol/l.

Outcomes and statistical analysis

The primary outcome was the time spent with plasma glucose levels in the hypoglycaemic range (<3.9 mmol/l) during exercise and the following 1 h recovery period. Secondary outcomes were: (1) decrease in plasma glucose levels during exercise; (2) number of participants experiencing at least one exercise-induced hypoglycaemia event requiring treatment; (3) total number of exercise-induced hypoglycaemia events; (4) during the exercise period only, the percentage of time spent with plasma glucose levels <3.9 mmol/l, <3.3 mmol/l, 3.9–7.8 mmol/l, 3.9–10 mmol/l, >10 mmol/l and >13.9 mmol/l, and mean plasma glucose at the start of exercise, plasma insulin concentration and insulin delivery rate, and SD and CV of glucose values. These glucose ranges were chosen in accordance with the recommendations of the consensus about outcome measures reported in artificial pancreas studies [25].

The main comparisons were made between unannounced exercise and each of the two announced strategies separately. The sample size calculation for the primary outcome was estimated by assuming a mean percentage of time spent with glucose levels <3.9 mmol/l of 6.84% (SD 8.35%) from a previous closed-loop study [6]. A sample size of 37 participants was needed to provide 80% power to detect a reduction of at least 5% in the primary outcome (Cohen’s medium effect size around 0.6) with announced strategies compared with the unannounced strategy.

Continuous variables were presented as mean±SD; medians with IQR (25th–75th percentile) were additionally reported when dealing with skewed data. Categorical variables were presented as frequencies and percentages. A multivariate linear mixed-effects model was used to compare continuous outcomes between different strategies, with strategy sequence, period and strategy type (fixed effect), and participant nested within sequence (random effect), entered as covariates. The effect of study strategies on the categorical outcomes was assessed using a random-effects generalised linear mixed model with logit link function for binary outcomes and log link function (assuming a Poisson distribution) for count outcomes. A bootstrap resampling method with no replacement using 500 samples was used to estimate the 95% CIs and p values, thus testing the model’s parameters. Using this approach, there was no need to check for distributional assumptions such as normality. Two-tailed p values <0.05 were considered statistically significant. All analyses were performed using SAS software, version 9.4 (SAS Institute, USA). All data were included in the analysis and no data imputation (last observation carried forward) or exclusion was performed after hypoglycaemia events.

Public and patient involvement

No members of the public or patients were involved in the design, conduct or interpretation of the study.

Results

Between September 2016 and November 2018, adults living with type 1 diabetes were enrolled at the Montreal Clinical Research Institute. In total, 39 adults with type 1 diabetes using CSII were recruited of whom 37 (20 women) completed the study. Two participants dropped out: one because of scheduling problems and one because of a knee injury not related to the study. The baseline characteristics of the participants are summarised in Table 1.

Table 1 Baseline characteristics of the 37 study participants (17 men and 20 women)

During exercise and the following 1 h recovery period, the mean time spent in hypoglycaemia (<3.9 mmol/l) was lower during the exercise announcement strategies (A-RB 2.0 ± 6.2% and A-FB 7.0 ± 12.6%) compared with the unannounced strategy (U-FB 13.0 ± 19.0%) (p < 0.0001 and p = 0.005, respectively) (Table 2, Fig. 1). The comparison between the two announced exercise strategies yielded a p value of 0.06. Time spent in hyperglycaemia (>10 mmol/l) for the A-RB strategy was increased by 24.6% in comparison with the A-FB strategy (p = 0.0001), and by 21.0% in comparison with the U-FB strategy (p = 0.001) (Table 2). Time spent in target (3.9–10 mmol/l) was decreased by 12.7% during A-RB in comparison with U-FB (p = 0.05) (Table 2).

Table 2 Summary and comparison of study outcomes for the three interventions
Fig. 1
figure 1

Individual data points for the study primary outcome of time spent in hypoglycaemia (plasma glucose [PG] <3.9 mmol/l)

Exercise announcement reduced the proportion of participants who experienced at least one hypoglycaemia event during exercise requiring treatment to 5.4% with A-RB, 10.8% with A-FB vs 16.2% with U-FB. A similar trend was seen in the following 1 h recovery period (Table 2). During exercise, none of the participants required repeated correction of hypoglycaemia with the A-RB strategy, while one participant during the A-FB and two participants during the U-FB strategies received repeated correction.

Participants began the exercise session with a higher mean plasma glucose for the A-RB strategy (12.8 ± 2.9 mmol/l) compared with both the A-FB strategy (11.4 ± 3.3 mmol/l; p = 0.009) and the U-FB strategy (11.4 ± 3.8 mmol/l; p = 0.008). Over the whole exercise period, mean plasma glucose was higher for the A-RB strategy in comparison with the A-FB (p = 0.01) and U-FB strategies (p = 0.001) (Table 2). The mean decrease in plasma glucose (glucose level at the end of exercise minus glucose level at the beginning of exercise) was lower with the A-RB strategy (−0.3 ± 2.8 mmol/l) compared with the A-FB (−2.6 ± 2.9 mmol/l; p < 0.0001) and U-FB (−2.4 ± 2.7 mmol/l; p < 0.0001) strategies.

When exercise was announced to the closed-loop system, plasma insulin concentrations over the whole exercise period were lower (median, IQR) for the A-RB (157.2, 104.1–206.9 pmol/l) and A-FB (179.1, 128.6–251.7 pmol/l) strategies than for the U-FB strategy (193.3, 136.5–256.2 pmol/l; p = 0.0008 and p = 0.09, respectively). Differences in the rate of insulin delivery among the strategies during the exercise period did not reach statistical significance (Table 2). The electronic supplementary material (ESM Table 1) shows CGM-based outcomes of percentages of time spent at different glucose levels and thresholds. ESM Fig.1 can be consulted for individual plasma glucose curves superimposed for the three strategies (one graph for each of the 37 participants).

Discussion

We conducted the first closed-loop system trial that directly compared announced with unannounced strategies to reduce hypoglycaemia risk during postprandial exercise in individuals with type 1 diabetes. Our findings suggest that the A-RB strategy reduced hypoglycaemia risk compared with the A-FB and U-FB strategies, albeit with increased time spent in hyperglycaemia.

Previous studies that compared CSII with closed-loop systems during exercise have shown that closed loop was associated with an increase in the time spent in range and a decrease in the number of hypoglycaemia events [3, 19, 24, 26,27,28]. However, none of these studies have challenged closed-loop systems with exercise taken in the postprandial state. Postprandial exercise combines multiple challenges for a closed-loop system [9, 29, 30]: the difficulty to control postprandial glucose excursions [12, 30] and the inability to reduce circulating insulin levels secondary to insulin meal bolus during exercise-induced hypoglycaemia.

Thus, this closed-loop study investigated strategies to achieve adequate plasma insulin levels for postprandial exercise by taking into account the basal infusion rate and the meal bolus. Basal insulin rate adjustment in a closed-loop system was made by setting higher glucose targets. Accumulating data in CSII suggest that the basal insulin rate needs to be reduced about 90 min before exercise. This strategy would allow enough time for plasma insulin levels to drop by the time of exercise start given the pharmacokinetics of currently used insulins [11, 13, 14, 17, 20, 31]. Bolus reduction obviously has to be done at the time of meal consumption. A 50% reduction in meal bolus was recommended for moderate-intensity, continuous exercise [10,11,12]. This 50% recommendation was nonetheless based on a study arm that did not include a change in basal insulin rate infusion [11]. In a closed-loop system setting that dynamically adjusts the insulin infusion rate, we assumed that a lower reduction in meal bolus would be sufficient and thus tested a reduction by one-third (33%) instead of one-half (50%).

The effect of these insulin adjustments in the announced strategies (A-RB and A-FB) may be observed in their corresponding glucose profiles, plasma insulin levels and rate of hypoglycaemia events. Exercise announcement resulted in higher mean glucose levels during the exercise session in comparison with the unannounced approach (Table 2, Fig. 2). However, glucose levels at the beginning of exercise, i.e. 90 min after the meal bolus, were only higher with the A-RB strategy and therefore appear to be mainly influenced by the bolus reduction. This is expected, as rapid insulin analogues peak at around 90 min. The insulin adjustments were reflected in the plasma insulin values (Fig. 3) of the different strategies. Previous studies of post-absorptive exercise and insulin adjustments have shown a modest exercise-related increase in circulating insulin, followed by a decline throughout the exercise session [31, 32]. In the current study, during the U-FB strategy, insulin concentration increased over the whole exercise period without any observed decline. This was due to the active meal insulin bolus and to higher insulin infusion rates in the U-FB arm. On the other hand, the lowest plasma insulin levels were seen with the A-RB strategy, and middle values with the A-FB strategy. Interestingly, some increases in plasma insulin concentrations towards the last 20 min in the A-RB strategy were observed, reflecting a closed-loop algorithm response to increased glucose values with this strategy. These observations will help to improve the design of the algorithms of closed-loop systems in relation to physical activity.

Fig. 2
figure 2

Plasma glucose concentrations over the course of the three interventions. Shaded area corresponds to the exercise session. Data are expressed as mean (SD)

Fig. 3
figure 3

Plasma insulin concentrations over the course of the three interventions. Shaded area corresponds to the exercise session. Data are expressed as median (IQR)

As a result of changes in plasma insulin levels, time spent in hypoglycaemia and the number of events requiring treatment for hypoglycaemia (using 50 ml 20% dextrose i.v. or carbohydrate intake) were minimised in A-RB and improved in A-FB in comparison with U-FB (control closed-loop arm). Our results highlight the limitations of available closed-loop systems using current CGM and CSII methods given the slow pharmacokinetics of insulin and the lag time of CGM readings when blood glucose declines rapidly during exercise [9, 33, 34]. A physiological lag time is needed to equilibrate glucose between the blood and the interstitial fluid compartments. If left without interference, adjustments of insulin infusion rates by the closed-loop algorithm based on changing CGM readings during exercise, as is the case with the U-FB strategy, carried higher risks of hypoglycaemia. CGM accuracy is key for optimal closed-loop algorithm operation; unfortunately, exercise is generally associated with glucose overestimation by CGM devices [34]. Such overestimation may lead to higher insulin infusion; even the A-RB strategy did not fully eliminate hypoglycaemia risk but significantly reduced it. Future CGM algorithms should reduce this overestimation by increasing sampling rates and adjusting readings according to trends in glucose changes or other techniques of glucose sensing than glucose oxidase [35].

Our results for the U-FB strategy agree with those of other studies that used exercise detection methods that proved to be insufficient to completely prevent hypoglycaemia during aerobic exercise [17]. The addition of carbohydrate snacks before or during exercise may be considered but can cause rebound hyperglycaemia. Diabetic individuals may also find it challenging to maintain or lose weight if frequently adding extra calories to their daily intake. There is insufficient data to guide the amount or timing of such snacks in the setting of postprandial exercise during closed-loop control.

During the A-RB strategy, a higher time in hyperglycaemia was noted in comparison with the full bolus strategies. Until further progress in insulin formulations and CGM technology is achieved, this mild hyperglycaemia during exercise may be an acceptable trade-off given the existing limitations of closed-loop components and the complexity of exercise-induced pathophysiological changes in type 1 diabetes. Other potential solutions could include testing a lower bolus reduction and/or faster acting new insulins and/or self-learning algorithms, because hypoglycaemia risk differs widely from one diabetic individual to another.

Multi-hormonal closed-loop systems with the addition of pramlintide and/or glucagon would be interesting options to investigate for postprandial exercise. Pramlintide is an analogue of amylin that is co-secreted with insulin to delay gastric emptying and prevent postprandial hyperglycaemia but is deficient in type 1 diabetes. Pramlintide has been shown to improve postprandial glucose control in dual-hormone closed-loop systems, but its efficacy has not been assessed during postprandial exercise [36, 37]. The beneficial effect of glucagon addition to closed-loop systems was shown during physical activity but has been mainly tested in post-absorptive exercise [6]. Both of these hormones might have a role in fine-tuning the balance between hyper- and hypoglycaemia in postprandial exercise.

The study had some limitations. It was conducted in a controlled setting with one testing per strategy; therefore, confirming the validity of the proposed approaches should be replicated in outpatient trials and during several exercise sessions. The closed-loop control was applied an hour before each intervention; one could argue that longer closed-loop use (≥24 h) might have brought participants to better control prior to the exercise intervention, but this would have made the study very burdensome to the participants (staying in the testing centre for a long time with an overnight stay). It is noteworthy that the participants were all using CSII and were relatively well controlled, with a mean HbA1c that was close to the recommended target (56 mmol/mol; 7.3%). Intravenous glucose was used to correct hypoglycaemia events because the use of a face mask might not reflect the use of oral glucose in real-life settings. The lowest recommended dose for hypoglycaemia correction is 15 g using oral glucose and 10 g using i.v. dextrose [38]; hence, we chose the latter in our study. The type of exercise and its duration, intensity and timing are all factors that significantly influence blood glucose, so the proposed strategies of this study and other approaches would need to be tested in other exercise settings. Nevertheless, we started by testing a frequently chosen type of exercise (continuous exercise at moderate intensity) in type 1 diabetes and we picked convenient timing for most individuals (around 90 min would be needed on average to avoid exercising on a full stomach, and/or to reach a sports facility). The proposed strategy (A-RB) should be tested in multiple situations to ensure its applicability. On the other hand, the main strength of our trial is that it explored a clinical need under controlled conditions with a large number of participants and comparison of three scenarios.

In conclusion, in the context of moderate-intensity, continuous postprandial exercise undertaken 90 min after a meal, A-RB was superior in reducing the time spent in hypoglycaemia compared with the two other strategies. This approach was, however, associated with some increase in the time spent in hyperglycaemia during the exercise session. In the future, we plan to test whether the amount of reduction in meal bolus should be guided by the pre-meal blood glucose level, to mitigate the increased hyperglycaemia seen in our study. Future more advanced algorithms and/or multi-hormonal approaches could also be required to achieve the goal of preventing hypoglycaemia without inducing transient hyperglycaemia.